|
//------------------------------------------------------------
// Copyright (c) Microsoft Corporation. All rights reserved.
//------------------------------------------------------------
namespace System.Runtime.Collections
{
using System;
using System.Collections;
using System.Threading;
using System.Diagnostics;
// This cache works like a MruCache, but operates loosely and without locks in the mainline path.
//
// It consists of three 'hoppers', which are Hashtables (chosen for their nice threading characteristics - reading
// doesn't require a lock). Items enter the cache in the second hopper. On lookups, cache hits result in the
// cache entry being promoted to the first hopper. When the first hopper is full, the third hopper is dropped,
// and the first and second hoppers are shifted down, leaving an empty first hopper. If the second hopper is
// full when a new cache entry is added, the third hopper is dropped, the second hopper is shifted down, and a
// new second hopper is slotted in to become the new item entrypoint.
//
// Items can only be added and looked up. There's no way to remove an item besides through attrition.
//
// This cache has a built-in concept of weakly-referenced items (which can be enabled or disabled in the
// constructor). It needs this concept since the caller of the cache can't remove dead cache items itself.
// A weak HopperCache will simply ignore dead entries.
//
// This structure allows cache lookups to be almost lock-free. The only time the first hopper is written to
// is when a cache entry is promoted. Promoting a cache entry is not critical - it's ok to skip a promotion.
// Only one promotion is allowed at a time. If a second is attempted, it is skipped. This allows promotions
// to be synchronized with just an Interlocked call.
//
// New cache entries go into the second hopper, which requires a lock, as does shifting the hoppers down.
//
// The hopperSize parameter determines the size of the first hopper. When it reaches this size, the hoppers
// are shifted. The second hopper is allowed to grow to twice this size. This is because it needs room to get
// new cache entries into the system, and the second hopper typically starts out 'full'. Entries are never added
// directly to the third hopper.
//
// It's a error on the part of the caller to add the same key to the cache again if it's already in the cache
// with a different value. The new value will not necessarily overwrite the old value.
//
// If a cache entry is about to be promoted from the third hopper, and in the mean time the third hopper has been
// shifted away, an intervening GetValue for the same key might return null, even though the item is still in
// the cache and a later GetValue might find it. So it's very important never to add the same key to the cache
// with two different values, even if GetValue returns null for the key in-between the first add and the second.
// (If this particular behavior is a problem, it may be possible to tighten up, but it's not necessary for the
// current use of HopperCache - UriPrefixTable.)
class HopperCache
{
readonly int hopperSize;
readonly bool weak;
Hashtable outstandingHopper;
Hashtable strongHopper;
Hashtable limitedHopper;
int promoting;
LastHolder mruEntry;
public HopperCache(int hopperSize, bool weak)
{
Fx.Assert(hopperSize > 0, "HopperCache hopperSize must be positive.");
this.hopperSize = hopperSize;
this.weak = weak;
this.outstandingHopper = new Hashtable(hopperSize * 2);
this.strongHopper = new Hashtable(hopperSize * 2);
this.limitedHopper = new Hashtable(hopperSize * 2);
}
// Calls to Add must be synchronized.
public void Add(object key, object value)
{
Fx.Assert(key != null, "HopperCache key cannot be null.");
Fx.Assert(value != null, "HopperCache value cannot be null.");
// Special-case DBNull since it can never be collected.
if (this.weak && !object.ReferenceEquals(value, DBNull.Value))
{
value = new WeakReference(value);
}
Fx.Assert(this.strongHopper.Count <= this.hopperSize * 2,
"HopperCache strongHopper is bigger than it's allowed to get.");
if (this.strongHopper.Count >= this.hopperSize * 2)
{
Hashtable recycled = this.limitedHopper;
recycled.Clear();
recycled.Add(key, value);
// The try/finally is here to make sure these happen without interruption.
try { } finally
{
this.limitedHopper = this.strongHopper;
this.strongHopper = recycled;
}
}
else
{
// We do nothing to prevent things from getting added multiple times. Also may be writing over
// a dead weak entry.
this.strongHopper[key] = value;
}
}
// Calls to GetValue do not need to be synchronized, but the object used to synchronize the Add calls
// must be passed in. It's sometimes used.
public object GetValue(object syncObject, object key)
{
Fx.Assert(key != null, "Can't look up a null key.");
WeakReference weakRef;
object value;
// The MruCache does this so we have to too.
LastHolder last = this.mruEntry;
if (last != null && key.Equals(last.Key))
{
if (this.weak && (weakRef = last.Value as WeakReference) != null)
{
value = weakRef.Target;
if (value != null)
{
return value;
}
this.mruEntry = null;
}
else
{
return last.Value;
}
}
// Try the first hopper.
object origValue = this.outstandingHopper[key];
value = this.weak && (weakRef = origValue as WeakReference) != null ? weakRef.Target : origValue;
if (value != null)
{
this.mruEntry = new LastHolder(key, origValue);
return value;
}
// Try the subsequent hoppers.
origValue = this.strongHopper[key];
value = this.weak && (weakRef = origValue as WeakReference) != null ? weakRef.Target : origValue;
if (value == null)
{
origValue = this.limitedHopper[key];
value = this.weak && (weakRef = origValue as WeakReference) != null ? weakRef.Target : origValue;
if (value == null)
{
// Still no value? It's not here.
return null;
}
}
this.mruEntry = new LastHolder(key, origValue);
// If we can get the promoting semaphore, move up to the outstanding hopper.
int wasPromoting = 1;
try
{
try { } finally
{
// This is effectively a lock, which is why it uses lock semantics. If the Interlocked call
// were 'lost', the cache wouldn't deadlock, but it would be permanently broken.
wasPromoting = Interlocked.CompareExchange(ref this.promoting, 1, 0);
}
// Only one thread can be inside this 'if' at a time.
if (wasPromoting == 0)
{
Fx.Assert(this.outstandingHopper.Count <= this.hopperSize,
"HopperCache outstandingHopper is bigger than it's allowed to get.");
if (this.outstandingHopper.Count >= this.hopperSize)
{
lock (syncObject)
{
Hashtable recycled = this.limitedHopper;
recycled.Clear();
recycled.Add(key, origValue);
// The try/finally is here to make sure these happen without interruption.
try { } finally
{
this.limitedHopper = this.strongHopper;
this.strongHopper = this.outstandingHopper;
this.outstandingHopper = recycled;
}
}
}
else
{
// It's easy for this to happen twice with the same key.
//
// It's important that no one else can be shifting the current oustandingHopper
// during this operation. We are only allowed to modify the *current* outstandingHopper
// while holding the pseudo-lock, which would be violated if it could be shifted out from
// under us (and potentially added to by Add in a ----).
this.outstandingHopper[key] = origValue;
}
}
}
finally
{
if (wasPromoting == 0)
{
this.promoting = 0;
}
}
return value;
}
class LastHolder
{
readonly object key;
readonly object value;
internal LastHolder(object key, object value)
{
this.key = key;
this.value = value;
}
internal object Key
{
get
{
return this.key;
}
}
internal object Value
{
get
{
return this.value;
}
}
}
}
}
|